Reinforcement Learning Framework for Tactical Decision-Making in Curling

A recent study published on arXiv (2608.02379) presents a reinforcement learning model aimed at enhancing tactical decision-making in curling, often referred to as "Chess on Ice" due to its intricate strategies. Unlike chess, curling has seen minimal exploration within machine learning, with earlier efforts primarily focused on statistical methods. The authors introduce a framework that quantitatively assesses and contrasts tactical choices. The sport poses challenges, including continuous state and action spaces, unpredictable outcomes influenced by player skill, and state transitions that are highly sensitive to minor changes. To tackle these issues, they utilize the Deep Deterministic Policy Gradient (DDPG) actor-critic algorithm, tailored for the game's finite-horizon nature. Their experiments reveal that effective curling strategies can be developed entirely through self-supervision, without the need for human-annotated data, as demonstrated with a simplified four-rock version. This research is noteworthy for applying sophisticated AI techniques to a sport that has been largely overlooked in machine learning, potentially providing fresh perspectives for sports analytics and decision-making.